Captions (green screen, voice-synced, Shorts-safe)
The user edits captions further in After Effects, so captions are NEVER burned into the film. Deliver a separate clip on pure green (#00FF00), same size, fps and frame count as the film, which they key with Keylight.
Look (agreed 2026-11-11, from the user's "LOCKED" reference frame)
- Position: centred horizontally; the middle of the capitals at 71.5 % of the frame height (under the character's chest). The middle of the capitals is the anchor, so a comma or a Q never moves a caption.
- White, UPPERCASE, Arial Black, 3.2 % of frame height (82 px at 2560). It was 4 % at first, and the user asked for smaller so the Shorts UI never covers it. Black outline; hard black drop shadow down-right. The shadow is solid on purpose: a blurred/semi-transparent shadow keys into a green-grey fringe.
- Punctuation is kept as in the script (MONA LISA. / PARIS, / 1911.). The user asked for it.
--strip-punctremoves trailing . , ; : - 2-3 words max per caption. A caption never crosses a sentence/comma or a pause in the voice, never ends on a glue word (the, of, with, yang, di, dengan ...), never splits a name (Mona | Lisa), avoids lone single words.
- Each caption appears on the frame its first word is spoken (a 5-frame pop 0.85 -> 1.06 -> 1) and holds until the next one. After a pause in the voice it leaves 0.25 s after its last word.
YouTube Shorts safe zone (must hold for every caption, at the pop's 1.06 peak)
Conservative overlay areas:
- action buttons: right 16 % of the width
- title/channel/music: bottom 25 % of the height
- top bar: top 8 % of the height
The script measures every caption's ink, prints WARNING outside the Shorts safe zone: <text> for any that cross a
zone, and prints the reach (should be about "83.8 % of width, 69.8-74.0 % of height"). Chunks are split so they fit
--max-width 0.68. A single long word is shrunk to fit. Never deliver with a WARNING: lower --size, or raise
--y a little.
Model (token saving, approved 2026-11-12)
Run captions on Haiku 5.5 (remotion-renderer agent with model: "haiku", or a Haiku session): the script does the
timing, splitting and safe-zone checks and prints them as text. Sonnet only to judge a new caption look; Opus only to
change scripts/captions.py logic.
Steps
- Word timings: the short's
stories/<slug>/voice/words.json. It is made byscripts/words_from_voice.py(faster-whisper aligned to the script's spelling). If it is missing, make it first:python scripts/words_from_voice.py <voice.wav> <SCRIPT.md> <words.json> [en|id]. The times must match the film's audio (the final's voice starts at 0 s). - Test span first (the user wants to SEE a short test before a full render):
It prints every caption with its times and the safe-zone check, and writes:python scripts/captions.py stories/<slug>/voice/words.json out/<slug>/captions --start 0 --dur 10 \ --preview "out/<slug>/final/<Title>.mp4" --alphacaptions_green.mp4captions_alpha.movcaptions.jsoncaptions_preview.mp4(film + keyed captions + sound, for checking only)
- Check:
- Pull frames of the widest captions from the preview with the Shorts zones drawn in red (ffmpeg
drawbox: x=iw0.84 from y=ih0.45, y=ih0.75 to the bottom, the top ih0.08),hstackthem into one sheet, and look. - Look at one full-res crop for green fringe.
- Send the user
captions_preview.mp4,captions_green.mp4and the printed caption list.
- Pull frames of the widest captions from the preview with the Shorts zones drawn in red (ffmpeg
- After the user's OK, render full length with
--dur <film frames / fps>so the clip has exactly the film's frame count (check withffprobe -count_packets).
Knobs (scripts/captions.py)
--y 0.715: position--size 0.032: font size--max-width 0.68: widest caption--font <ttf>--max-words 3--gap 0.30: pause that breaks a phrase--hold 0.25--lead 0: frames early--no-pop--lower: keep case--strip-punct--alpha: ProRes 4444
AE notes for the user
- Green clip: use Keylight with Screen Colour = pure green. It is yuv444 crf 8, so edges stay clean.
- Alpha .mov: needs no key at all (drop it on top).
- captions.json: lists each caption's text and in/out seconds, for rebuilding them as text layers.